Primary sourcearXiv cs.LG
View sourceEvent ·Multi-source
9个LLM结构化输出失败实证研究:Markdown围栏是最大坑
研究覆盖 9 个 OpenAI 和 Anthropic 模型/API 配置,共 3,997 次生成尝试,任务来自 HotpotQA 和 FinanceBench。作者提出细粒度的 schema 失败分类法和确定性分类器,发现相同的有效率可能掩盖不同的失败模式。Markdown 围栏包裹的 JSON 是最常见失败:681 个围栏失败中仅 48.0% 在去围栏后能通过校验,其余多隐藏字段长度违规。研究总结了 4 条工程实践,涉及校验器、schema、重试和评估设计,并开源了分类器、补全记录和标注数据。
Current read
如果你的 Agent 管线老被 Pydantic 校验卡住,这篇拿 9 个模型、近 4000 次调用跑出来的失败分析值得看看。
11 sources36° AI heatLast updated 2026/10/24 23:30:38
Evidence
11 sourcesRelated source 1TestingCatalog
View sourceRelated source 2Teortaxes
View sourceRelated source 3Nathan Lambert
View sourceRelated source 4DAIR.AI
View sourceRelated source 5IT之家
View sourceRelated source 6Blognone
View sourceRelated source 7Anthropic: Research
View sourceRelated source 8Ethan Mollick
View sourceRelated source 9marktechpost
View sourceRelated source 10VnExpress Số hóa
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.